Using Contrastive Learning for Aspect Detection with Vietnamese Dataset

Abstract

Customer reviews for a business are extremely important because they provide valuable insights, and impact the customer’s decision-making process and the customer experience after a purchase. Determining the aspect of the reviews, which makes business easier to analyze, is a difficult problem because each review has a different writing style, many grammatical errors, and often contains acronyms. Unsupervised aspect detection (UAD) strives to automatically extract understandable facets and pinpoint segments that are specific to these facets from online reviews. However, because of the difference between syllables and accents in Vietnamese, automatically extracting aspects in reviews is a challenging task. To address aspect detection issues in the context of Vietnamese text, in this thesis, we will propose an approach that combines contrastive learning with aspect detection. Specifically, we generate aspects for similar word clusters followed by the model is encouraged to differentiate between them and capture distinctive features by measuring the similarity between pairs of samples in the Vietnamese dataset. This enables the model to generate high-quality representations for aspects and their corresponding text segments. Furthermore, this thesis builds upon the experimental methods used by previous studies such as Smooth self-attention (SSA) and High-resolution selective mapping (HRSMap) to further enhance the performance of aspect detection in the context of Vietnamese text. We achieved relatively good results for the 4 "golden" aspects, averaging around 0.73 for F1-score